Abstract
This doctoral research focuses on developing and applying data-driven activity classification tools based on machine-learning algorithms, as well as physically motivated simulated spectral energy distribution (SED) libraries that characterize stellar populations and dust properties across diverse star-forming environments.We developed robust activity diagnostic tools that minimize the need for data preprocessing for their application. The first tool employs three easily measured features, the equivalent widths (EW) of three prominent spectral features: Hβ, [O III]λ5007, and Hα + [N II] λλ6548,84. The classifier is based on a support vector machine (SVM) trained on the EW of these spectral lines measured for optically classified galaxies using other well-established methods (ground-truth). The tool achieves ∼83% overall accuracy and with completeness per class (recall) of 79% for star-forming, 94% for active galactic nucleus (AGN), 85% for low-ionization nuclear emission-line regions (L ...
This doctoral research focuses on developing and applying data-driven activity classification tools based on machine-learning algorithms, as well as physically motivated simulated spectral energy distribution (SED) libraries that characterize stellar populations and dust properties across diverse star-forming environments.We developed robust activity diagnostic tools that minimize the need for data preprocessing for their application. The first tool employs three easily measured features, the equivalent widths (EW) of three prominent spectral features: Hβ, [O III]λ5007, and Hα + [N II] λλ6548,84. The classifier is based on a support vector machine (SVM) trained on the EW of these spectral lines measured for optically classified galaxies using other well-established methods (ground-truth). The tool achieves ∼83% overall accuracy and with completeness per class (recall) of 79% for star-forming, 94% for active galactic nucleus (AGN), 85% for low-ionization nuclear emission-line regions (LINERs), 77% for composite, and 96% for passive galaxies. The main advantages of this tool are: (a) it can be used on objects exhibiting emission or absorption lines (thus seamlessly including passive and active galaxies), it does not require any data processing, including starlight subtraction, modeling or deblending of the spectral lines or even flux calibration. A second classifier focusing on the identification of main-sequence star-forming galaxies uses custom-defined mid-IR polycyclic aromatic hydrocarbons (PAH) features (peaking at 6.2, 7.7 + 8.6, and 11.3 μm) and the dust continuum. This classifier correctly identifies 91% of main-sequence star-forming galaxies with low AGN contamination (∼16% misclassified as star- forming in our test sample). This classifier can be used for the identification of main-sequence star-forming galaxies in wide-area photometric or spectroscopic surveys. For this reason, we also provide a version involving the F560W, F770W, F1130W, and F2100W JWST/MIRI filters. This diagnostic tool is based on a physically motivated SED library built by simulating main-sequence galaxies based on star formation histories and probability distribution functions for dust parameters reported in the literature for a large sample of main-sequence star- forming galaxies. To extend our SED library to encompass star-forming environments with diverse physical conditions, we perform broadband SED fitting on representative samples of metal-poor dwarf galaxies, intermediate galaxies, and luminous infrared galaxies. From the derived best-fit parameters, we construct a simulation-based SED library by generating SEDs using parameter sets drawn from the posterior distribution of each galaxy, thereby preserving covariances between the stellar and dust parameters. The simulated library, containing nearly one million SEDs (∼18,000 metal-poor starbursts, ∼117,000 luminous infrared galaxies, ∼108,000 intermediate galaxies, and ∼797,000 main-sequence galaxies), spans four orders of magnitude in stellar mass (8 < log10 M⋆ < 12) and a wide star formation rate range (−2.5 < log10 SFR < 2.5). The analysis of the results of these SED fits showed that the frame work reliably separates bursty and main–sequence systems, with the ratio of stellar population age to the star-formation timescale (tage/τ ) serving as a stable diagnostic despite age- metallicity-dust degeneracies. Low-metallicity galaxies require power-law dust models rather than standard templates, and merging systems tend to underestimate stellar mass while preserving SFR estimates. A third project focused on the composition and ionization state of the dust in star-forming and AGN host galaxy. We used Spitzer/IRS spectra to isolate the PAH emission spectrum of galaxies using spectral decomposition tools (PAHFIT). Then, using the NASA Ames PAH IR Spectroscopic Database (PAHdb) we recovered the intrinsic PAH properties (such as ionization fraction and molecular sizes). The inferred PAH properties, along with properties of the host galaxy (e.g., star formation rate), serve as target variables, which are predicted from the galaxy’s PAH spectrum using a regression forest model, producing a fast, reliable probe of PAH chemistry and interstellar medium (ISM) conditions. Our analysis shows systematically larger PAHs in star-formation-dominated systems than in AGN hosts, while optically identified AGN show a broader range of ionization fractions, a subpopulation (mainly Seyfert 2) with PAH signatures and [Ne II] 12.81 μm emission consistent with circumnuclear star formation. Taken together, the PAH-based diagnostic for main-sequence star-forming galaxies and the spectral regression tool, the versatile activity classifier, and the physically motivated SED library of diverse star-forming environments form a coherent toolbox for probing galaxy conditions: the library provides realistic, diverse spectra for forward modeling and evolution studies; our regressor links mid-infrared features to conditions of the ISM (ionization state and molecule size); and the diagnostics enable automated, fast, and widely applicable identification of galaxy activity with minimal preprocessing. These methods are particularly well suited to leverage modern infrared facilities (e.g., JWST) and large spectroscopic surveys, facilitating population-level studies of ISM chemistry, star formation, and nuclear activity across cosmic environments.
show more
